AgenticInterpBench (HyVE) - Task Accuracy: leaderboard
Circuit-explanation benchmark, task level: the share of circuits whose derived task description an LLM judge (the lower of GPT-5.4 and Gemini-3.1-Pro) accepts, in percent, higher is better. 84 semi-synthetic InterpBench transformers with known circuits (163 annotated components); the model is the backbone of the authors' HyVE explainer agent, which observes activations, tests hypotheses with code and explains each component. Subjects are compiled circuits, where scores can partly reflect informed guessing. Khan et al., 2026.
Metric: Task accuracy (%; LLM-judged, lower of two judges). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 83 | #91 |
| 2 | Claude Sonnet 4.6 | 75 | #158 |
| 3 | GPT-5.4 | 63 | #121 |
| 4 | Qwen 3 Coder 30B A3B Instruct | 25 | #973 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agenticinterpbench-hyve-task-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.